[Paper Review] ArcticNet: A Deep Learning Solution to Classify Arctic Wetlands

Ziyu Jiang, K. Von Ness+2 more|arXiv (Cornell University)|Jun 1, 2019

Underwater Acoustics ResearchEarth and Planetary Sciences50 references3 citations

TL;DR

ArcticNet is a deep learning model that classifies Arctic wetlands using high-resolution multispectral satellite imagery (Planet Dove CubeSats) and Digital Elevation Models (ArcticDEM), fusing RGB, NIR, and DEM data through late fusion to improve semantic segmentation. It achieves 93.12% accuracy on a novel 50 km² Arctic Wetland Dataset (AWD), outperforming single-modality models in distinguishing ecologically critical wetland types like channel fen and peat bog.

ABSTRACT

Arctic environments are rapidly changing under the warming climate. Of particular interest are wetlands, a type of ecosystem that constitutes the most effective terrestrial long-term carbon store. As permafrost thaws, the carbon that was locked in these wetland soils for millennia becomes available for aerobic and anaerobic decomposition, which releases CO2 and CH4, respectively, back to the atmosphere.As CO2 and CH4 are potent greenhouse gases, this transfer of carbon from the land to the atmosphere further contributes to global warming, thereby increasing the rate of permafrost degradation in a positive feedback loop. Therefore, monitoring Arctic wetland health and dynamics is a key scientific task that is also of importance for policy. However, the identification and delineation of these important wetland ecosystems, remain incomplete and often inaccurate. Mapping the extent of Arctic wetlands remains a challenge for the scientific community. Conventional, coarser remote sensing methods are inadequate at distinguishing the diverse and micro-topographically complex non-vascular vegetation that characterize Arctic wetlands, presenting the need for better identification methods. To tackle this challenging problem, we constructed and annotated the first-of-its-kind Arctic Wetland Dataset (AWD). Based on that, we present ArcticNet, a deep neural network that exploits the multi-spectral, high-resolution imagery captured from nanosatellites (Planet Dove CubeSats) with additional DEM from the ArcticDEM project, to semantically label a Arctic study area into six types, in which three Arctic wetland functional types are included. We present multi-fold efforts to handle the arising challenges, including class imbalance, and the choice of fusion strategies. Preliminary results endorse the high promise of ArcticNet, achieving 93.12% in labelling a hold-out set of regions in our Arctic study area.

Motivation & Objective

• To address the lack of reliable, high-resolution wetland mapping in Arctic regions due to complex micro-topography and non-vascular vegetation.
• To construct the first-of-its-kind Arctic Wetland Dataset (AWD) with 500 annotated 30m×30m patches across six wetland classes.
• To develop a deep neural network, ArcticNet, that fuses multi-modal data (RGB, NIR, DEM, NDVI) for improved semantic segmentation of Arctic wetlands.
• To mitigate class imbalance in wetland data through a novel data augmentation strategy.
• To evaluate model performance across different fusion strategies and post-processing techniques like CRF for spatial consistency.

Proposed method

• The model uses a patch-based, multi-modal encoder architecture with separate branches for RGB, NIR/DEM/NDVI inputs, followed by late fusion of feature maps for classification.
• Normalized Difference Vegetation Index (NDVI) is computed via Eqn. 1 to enhance vegetation health representation in the input modality.
• A data augmentation strategy is designed to balance the class distribution, particularly for underrepresented classes like channel fen and peat bog.
• Late fusion is employed to combine features from different modalities, improving discrimination over single-modality approaches.
• Conditional Random Fields (CRF) are applied as post-processing to refine segmentation maps and suppress isolated misclassified pixels.
• The full study area is segmented via sliding window inference, with results stitched into a complete semantic label map.

Experimental results

Research questions

• RQ1 Can deep learning models effectively classify Arctic wetland functional types using high-resolution, multi-spectral satellite imagery and DEM data?
• RQ2 How does late fusion of RGB, NIR, and DEM modalities improve classification accuracy compared to single-modality models?
• RQ3 To what extent can data augmentation strategies mitigate class imbalance in Arctic wetland datasets?
• RQ4 How do post-processing techniques like CRF enhance spatial consistency and reduce noise in semantic segmentation outputs?
• RQ5 How generalizable is the ArcticNet model to other Arctic regions, and what role do different spectral band combinations play in distinguishing wetland types?

Key findings

• ArcticNet achieved a test accuracy of 93.12% on a hold-out set of the Arctic Wetland Dataset, demonstrating strong performance in classifying six wetland types.
• The late fusion model significantly outperformed the RGB-only model, reducing misclassification of lakes as channel fen and improving overall segmentation fidelity.
• The RGB-only model overestimated channel fen by a large margin (likely due to inability to detect moisture and biomass via visible light alone), while the late fusion model reduced this error.
• CRF post-processing effectively suppressed isolated misclassified pixels, such as erroneous wetland labels in lakes, improving spatial coherence.
• Qualitative validation with drone imagery confirmed that the model correctly distinguished key wetland features, such as sedge-dominated channel fens and moss-covered peat bogs.
• Despite limitations in classifying non-wetland features like roads (misclassified as water or bog), the overall model outputs were deemed interpretable and reasonable by geospatial experts.